One voice helps you develop an idea. Name the edge you are working on and the questioner — the stone, since a whetstone is what it is for — asks back, against arguments in design that were never settled: whether a photograph is evidence or construction; whether the designer is an author. The other stress-tests an idea, pointing to where a text quietly decided for you.
An instruction to a model is not a rule. A paragraph in the prompt told the model not to invent a premise, and it invented them for six versions. So the check sits outside the model — a short list of marks anybody can read, which means it misses things and says which. A paraphrase is already an interpretation, so retrieval matches your literal words and refuses embeddings, at a cost in recall.
The name doubles zeti, from the Greek zetetic — "proceeding by inquiry" — with an echo of the Sanskrit neti neti, "not this, not this". What the two voices are →
Say what you are working on and where it resists you. The questions come back in your own words, against tensions in the field that were never settled. Nothing is saved here — the conversation is yours and this holds no copy of it, so download it if you want to keep it.
Pick up a transcript you saved earlier. It is read here in your browser and never sent anywhere — the conversation simply carries on from where you left it.
A glossary or briefing about a field you don't work in yet. Six lines of questioning across three sittings, asking where you stand — then it stops and the enquiry is yours. Read here in your browser, never uploaded.
A dialogue you saved, marked wherever the questioning failed you and with what it
should have asked instead. Mark it in your own editor — there is nothing to type here. Send the
.md; it is the file you already have and nothing else holds a copy of it.
You can take any of them back at any time, and that is not a negotiation.
Nothing here is scored, because a score would be about you and what gets made is the conversation. This shows how the goal was re-drawn as it went. It is private to you, it exists nowhere else, and it goes when the tab does.
zetizeti is a whetstone. A whetstone does not cut for you; it is the stone you draw your own edge against. It asks; it does not answer. It has two voices — one helps you develop an idea, the other stress-tests an idea to see if it holds; both ask questions in your own words, and neither hands you a conclusion.
The conversation is not a route to something else. A tool that teaches you something has a thing to deliver — a method, a correction — and the talking is how it travels; once it has arrived, the talking can be thrown away. Nothing is being delivered here. No answer is handed over because there is none to hand over.
What you leave with is the conversation. It did not exist before and now it does; nobody else holds a copy; and when the tab closes it is gone unless you take it. That is why nothing is kept here. It belongs to whoever was in it, and the download is how it survives the tab closing.
You name an edge — what you are trying to do, and where it resists you. The stone asks back, each question grounded in a real tension in the field that was never settled. It does not resolve the tension for you; it keeps you in the question long enough to think. This is the older harm it was built against: the answer dropped in too soon, before you have done the thinking that would make it yours. An information gap a sentence can close; an understanding gap only your own movement can close. The stone refuses to close the second one for you.
Bring an idea you want to pressure-test — a conclusion you reached in an enquiry, an answer a machine handed you, a claim you found anywhere — and paste it in. The stone questions that text instead of you. A settled idea, wherever it came from, is a conclusion already sitting in your understanding; the question worth asking is whether you reached it, or it was simply deposited there. So the stone does to the text what it does to your own premature certainty — it locates the spots where the text quietly decided something for you (called a thing "intuitive", "best practice", "obviously" — a judgement worn as a description) and asks you about those spots.
What this voice wants. Not to tell you the idea is wrong. It never grades the text, never marks it right or wrong — that would only be a second deposited verdict, this time the tool's. It wants to give you back the act of judging. It locates; you decide. It is the neti neti of the name turned outward — "not this, not this" — subtracting the text's smuggled certainties until what is left is yours to weigh.
Nothing is kept, and the reason is not privacy. A reading of a text is a frame somebody put on it once, and an average of such frames is a number about nothing. So a critique lives only while this tab is open, and there is no list, no history and no tally — not withheld from you, but never made.
When you paste a text in, nothing about it is decided by a model. It is first broken into segments, and each segment is tagged by plain, stated rules — is this describing, or is it quietly deciding? A value worn as a property (the clean interface), a directive (you should…), a verdict relayed as settled (obviously, best practice) — each is caught by a rule about grammar and a short list of words, not by anything trained on the open web. A second step, also only arithmetic, settles which of those spots the question will point at. Only then does the model arrive, and only to do one thing: phrase a single question about that one spot, in the text's own words. A last check confirms what comes back is still a question, and carries no verdict of its own.
This is deliberate, and it has three consequences worth naming plainly. The reading cannot fail — there is no model call to time out or hallucinate; the worst case is that a spot reads as neutral. It is reproducible — the same text yields the same spots every time, not a different mood on a different day. And it is inspectable — every spot it marks carries the reason it was marked, down to the exact word and the rule that caught it. You can ask the tool why, and get an answer you could check by hand.
Its blind spots are a list, not a mystery. What the tool treats as a deciding word is a short, readable list — not weights you must take on trust. It is deliberately blunt: it would rather miss a smuggled verdict than invent one where there is none. And when it does miss — and it will — the fix is a line added to that list, something you can see and argue with, not a retraining no one can audit. The model does the language; the rules do the locating; you do the judging. Where another tool asks you to trust what it cannot show you, this one shows you how little it is deciding.
An instruction to a model is not a rule. From v0.11.2 to v0.17.0 a paragraph in the system prompt told the model not to invent a premise, and it invented them the whole time — addressed to the thing that was doing it. A rule handed to the party it binds is a request, and it reads exactly like a rule in the source.
So the check sits outside the model, and it is deliberately shallow. A question is written, held, and read against a finite list of marks — a missing question mark, a set of forbidden constructions — regenerated once if it carries one, and only then released. Two things follow and both are worth having in front of you. It cannot arrive word by word, because the whole of it must exist before any of it can be read, and that is what the refusal costs. And the list is the limit of the claim. A fluent, well-formed question with none of those marks in it can still put a verdict in front of you, and nothing here would catch it; the audit says so in its own summary. What is enforced is what could be named in advance, which is less than the promise a tool would normally make here.
A paraphrase is already an interpretation. The moment anything restates what you said, it has decided what you meant. Retrieval here therefore matches your literal words — exact-word, unstemmed, no embeddings and no semantic search anywhere in the path. An embedding would find more; it loses recall, and the loss is the price.
A citation is a claim about the world, and no model may certify one. A wrong source in front of somebody thinking sends them somewhere wrong, and a model has no way to know it is wrong. So: the corpus is written rather than scraped, and no entry ships until its citations have been checked against real sources.
A separate question is whether the framing around those sources is any good, and that is not a matter of fact at all. It is a judgement, so a person makes it. Every entry carries a second flag saying whether anybody has read it that way yet, and only a person can clear one. While it asks, a panel beside the question shows the entries it drew on and the state each is in — including the many nobody has read yet.
If a larger model asked a better question, the question was doing the wrong work. Composing a sentence is language work. Deciding what is worth asking is judgement work, and the corpus holds that. The model here is therefore a cheap one, and that is a standing test rather than an economy: swap it upward and the sentences get better, and if the questions get better as well, then judgement had been leaking into the half that only writes. ⚠️ That swap has not been run on the current build.
The questioning discipline is Clean Language, a method built in psychotherapy for asking about somebody’s own words without substituting your own. It does not transfer whole, and that was measured rather than assumed: a test harness put roughly 1,900 question-and-reply pairs through it, and the classic felt-sense moves came back refused between a quarter and two-fifths of the time — asked of an object rather than a feeling, what kind of X is that X produces nonsense. The repertoire was cut to what survived, which turned out to be the moves treating what you brought as a proposition. The other borrowing is Freire’s: that a settled conclusion has been deposited in you rather than reached. He was describing a teacher and a classroom. The depositor here is a machine, and it deposits in answer to a question you asked it.
A dialogue asks no credential of either side. Two people make a good one without being licensed to; somebody with the wrong training makes a better one than an expert having a bad day. No body certifies it. Nobody is professionally good at it, because a conversation that ran the same way every time would stop being one.
That is why what gets made here cannot be graded. A grade needs a standard, a standard needs somebody entitled to set it, and for this there is no such person — which is a fact about conversations rather than a policy set here.
Questions with nothing behind them run out in four turns and turn into pleasantries. So each one is grounded in a tension practitioners argue about and have not settled — written out, with the register a person actually speaks it in before they have the term for it, the ways it usually fails, and the sources. Each entry says on its own face which state it is in: the citations under it are checked before it can ship, and whether a person has read the framing is a separate flag that only a person can clear.
This material is not the subject being taught. Nothing here is trying to get it into you, and the test for adding more is whether it gives a conversation more to be about — never whether it covers more ground. That is also why none of it is scraped: it is written and then checked, because a wrong source in front of somebody thinking sends them somewhere wrong.
The second voice is not a looser version of the first — it is the same discipline, held tighter. Both speak in questions and nothing else. Both reuse your exact words rather than paraphrasing them. Neither explains, advises, reassures, concludes, or scores. There are no grades here, no leaderboards, no percent-complete. Those measure a person. This makes a conversation, and you keep it. It asks the questions; you do the thinking — whether a premature answer came from you, or from a machine.
It asks a great deal of patience. People who do not have it to give leave, and most people do not have it to give.
The method core, Part A, is one note where it should be many. Until v0.16.0 the second voice asked two-box questions: is it this, or is it that. A student picked from those boxes instead of describing what was in front of them. It ran that way for months. It was found because a student said it felt off and could not say why.
All of this is open and checkable — the code, the corpus, the locating rules, and the full verification records (citation ledgers, adversarial passes, sign-off sheets): github.com/zetizeti/zetizeti (AGPL-3.0). Nothing here asks to be trusted; it asks to be read.
The framework underneath is split-domain cognition. Language work and judgement work are two unlike kinds of thinking, and a tool that runs both through one channel ends up doing the judging for you. So the model here only phrases the question. The locating, the guard and the sign-off are code, or a person.
zetizeti was built alongside Koher, a ten-year practice of free open-source tools. Koher holds that a tool separating language from judgement has to show its architecture, and the curtain on this page is that habit: the retrieved tensions, the sources, and whether a person has read each one yet. The engine and the corpus are zetizeti's own.
Made by Prayas Abhinav.
Bring an idea you want to stress-test — a conclusion you reached in an enquiry, an answer a machine gave you, a claim from anywhere — and the stone questions the text to find out whether it holds: pressing where it might not, pointing to where it quietly decided something, where a judgement was slipped in as fact. It never grades it, never says "this is wrong": it locates, you judge. Each critique lives only in this tab — nothing is saved, and it is never scored, ranked, or compared. Each stands alone, then it is gone.
A specification is a description of a thing that does not exist yet, in your own words, complete enough that somebody or something else could build it. Bring one you are writing — for a sketch, a script, an object, a service — and the stone asks along the eight lines a specification is written in: what the thing is, what it holds, what changes, what decides, what happens at the edges, which values are fixed or free, how anybody would know it works, and what it will not know. Your answers go into the specification in your own words, when you say so. It never tells you what is missing and never completes it for you: whatever you have not specified will be decided by something else, and the point is that you find out what. Nothing is saved, nothing is scored.
Key usage across users — pool and personal — with turns, spend, and when each was last active. The pool figure runs against the ₹ ceiling; personal spend is shown apart. Read-only; no message content is shown. Figures are the real billed cost, converted at the live rate.